The good way to assess the independence of the sensor signals is the application of the statistical analysis of the variance
of these signals. We have applied the ANalysis Of VAriance (ANOVA) analysis [18] of the measured sensor signals for all types of coffee. In each case, we have got a p-value around 10−13 (probability that the sensor signals come from the same population—the so-called null hypothesis). This small p-value means that the null hypothesis should be rejected.
Fig. 4
presents the box plots produced by ANOVA corresponding to the original and normalized sensor signals for arabica. Very
similar plots have been obtained for other types of coffee mixtures.
Each box in Fig. 4 has lines at the lower quartile, median, and upper quartile values. Whiskers extend from each end of
the box to the adjacent values in the data. By default, the most extreme values within 1.5 times the interquartile range from the ends of the box. The points displayed with a red + sign denote outliers, i.e., the data with values beyond the ends of the whiskers.
In both cases (the original and normalized), the signals of the sensors occupy different ranges of values and have different
values of median, and we observe a lot of outliers. This is also the additional evidence of statistical independence of the sensor signals.
It is interesting to present the distribution of sensor signals on the radar plot. To present them in a more visible and understandable way, we have limited the presentation to only the five most representative classes of coffee blends: pure arabica A, pure robusta R, and three different mixtures RA10, RA50, and RA90.
Only one case of the measurement of each blend corresponding to the last pattern of sensor signals within the measurement window is depicted in Fig. 5. This plot confirms the variety of sensor responses corresponding to different mixtures of robusta and arabica. It is seen that the patterns of responses of the sensors are changing nonuniformly and the nonlinear methods of signal processing are needed to associate each pattern with a
particular type of blend.
The proper set of sensors used in the array should generate the independent signals for two different types of coffee. We
have calculated the cross-correlation coefficients of the signals obtained at 100 measured points for all 12 sensors, at the recognition of pure robusta and arabica. The selected results of the calculations concerning robusta and arabica are presented in Table I. They confirm relatively weak correlation existing between signals corresponding to these two types of coffee.